Downloaded from https://www.everyspec.com ADS-79D-HDBK All existing data that provides sensor data responding to both normal operation and failure conditions should be consolidated in a data warehouse for use in algorithm development. Assessing the data to determine data ―gaps‖ can provide insight into any additional testing or modeling and simulation required to support algorithm development. Performance metrics for the Diagnostic and Prognostic modules should be established for use in the validation and verification of the diagnostic and prognostic algorithms and the maintenance actions and maintenance credits which result.
Since the mathematical processes produce results which are estimates of the probability of the existence of faults and RUL, CIs and RUL confidence levels should be established. For CIs this is commonly expressed as a false positive rate, such as 10% false positives (detecting the existence of a fault that is not present). The Diagnostic Module should deliver results that provide determination with high confidence of the following characteristics: a. Accuracy: The proportion of all healthy and faulted components which were diagnosed correctly. Accuracy represents the most fundamental metric of an algorithm‘s performance. [Reference c] b. Detectability:
The extent to which a diagnostic measure is sensitive to the presence of a particular fault. Detectability should relate the smallest fault signature that can be identified at the prescribed false positive rate. c. Identifiability : The extent to which a diagnostic measure distinguishes one fault from another that may have similar properties. d. Separability: The extent to which a diagnostic measure discriminates between faulted and healthy populations. Any development of CIs for use in diagnostics should include the metrics above and a validation of those metrics. Only those CIs capable of being detected with high confidence should be used in deployed CBM systems.
Algorithms used to preprocess the sensor data (de-noising, filtering, synchronous time averaging (STA)) compress and reduce the data necessary to extract or develop the feature or CI used to confirm the presence of a fault. The preprocessing routines, selected for the application, are intended to improve the signal to noise ratio to correspondingly improve the probability of fault detection. Best practice and experience for the specific application may develop guidelines regarding the best range of signal to noise ratio for feature extraction. If those guidelines exist, every effort should be made to develop algorithms consistent with best practice.
The sub-process labeled Detection Algorithm Development (DAD) is often an iterative process that optimizes the data compression filtering and de-noising steps to develop the most effective group of features/CIs to be used as inputs to the diagnostic process. That process can create a feature ―vector‖ or group of individual features/CIs to be used to provide the most effective inputs to the diagnostic process. Data from actual failures or seeded fault testing, along with confirmation gained from Inspection/Tear Down Analysis (I/TDA), is used to evaluate the features and optimize their use for diagnosis.
The algorithms that calculate each CI can also evaluate the value of the CI against values or ―thresholds‖ that define the fault severity. An individual CI can be assigned values that are ―normal‖, ―marginal‖ (indicating potential for action such as ordering a part or scheduling a maintenance task), or ―abnormal‖ 74